Artificial Intelligence In Medicine 119 (2021) 102157

Contents lists available at ScienceDirect

Artificial Intelligence In Medicine

journal homepage: www.elsevier.com/locate/artmed

Dengue models based on machine learning techniques: A systematic

literature review
William Hoyos a, b, *, Jose Aguilar b, c, d, Mauricio Toro b
a
Grupo de Investigaciones Microbiológicas y Biomédicas de Córdoba, Universidad de Córdoba, Montería, Colombia
b
Grupo de Investigación en I+D+i en TIC, Universidad EAFIT, Medellín, Colombia
c
Centro de Estudios en Microelectrónica y Sistemas Distribuidos, Universidad de Los Andes, Mérida, Venezuela
d
Universidad de Alcalá, Depto. de Automática, Alcalá de Henares, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: Background: Dengue modeling is a research topic that has increased in recent years. Early prediction and
Dengue decision-making are key factors to control dengue. This Systematic Literature Review (SLR) analyzes three
Diagnostic model modeling approaches of dengue: diagnostic, epidemic, intervention. These approaches require models of pre­
Epidemic model
diction, prescription and optimization. This SLR establishes the state-of-the-art in dengue modeling, using ma­
Intervention model
Machine learning
chine learning, in the last years.
Methods: Several databases were selected to search the articles. The selection was made based on Preferred
Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology. Sixty-four articles were
obtained and analyzed to describe their strengths and limitations. Finally, challenges and opportunities for
research on machine-learning for dengue modeling were identified.
Results: Logistic regression was the most used modeling approach for the diagnosis of dengue (59.1%). The
analysis of the epidemic approach showed that linear regression (17.4%) is the most used technique within the
spatial analysis. Finally, the most used intervention modeling is General Linear Model with 70%.
Conclusions: We conclude that cause-effect models may improve diagnosis and understanding of dengue. Models
that manage uncertainty can also be helpful, because of low data-quality in healthcare. Finally, decentralization
of data, using federated learning, may decrease computational costs and allow model building without
compromising data security.

1. Introduction
Dengue is a vector-borne disease, with high importance in public
health [1]. This disease is widely distributed worldwide; especially, in
tropical and subtropical areas [2]. The disease is produced by an arbo­
virus (DENV) that receives the same name. To date, four virus serotypes
have been identified: DENV-1, DENV-2, DENV-3 and DENV-4 [3]. The
infection is transmitted to humans by the bite of mosquitoes of the genus
Aedes, mainly A. aegypti and A. albopictus [4].
In 1997, the World Health Organization (WHO) classified the disease
like dengue fever and dengue hemorrhagic fever [5]. A new classifica­
tion was proposed in 2009, which was based on the severity level of the

Abbreviations: DENV, dengue virus; WHO, World Health Organization; SD, severe dengue; DSS, dengue shock syndrome; SLR, systematic literature review;
PRISMA, preferred reporting items for systematic reviews and meta-analyses; LoR, logistic regression; ANN, artificial neural networks; MLP, multilayer perceptron;
MCS, modified cuckoo search algorithm; PSO, particle swarm optimization; SE, structural equations; AUC, area under the curve; OR, odds ratio; CI, confidence
intervals; RF, random forest; PAF, platelet activating factor; S1P, sphingosine 1-phosphate; IL-1β, interleukin 1 beta; TNFα, tumor necrosis factor alpha; IL-10,
interleukin 10; FL, fuzzy logic; VEGF, vascular endothelial growth factor; SVM, support vector machines; DNA, deoxyribonucleic acid; DT, decision trees; APRI,
aspartate aminotransferase/platelet count ratio index; LASSO, adaptive least absolute shrinkage and selection operator; GAM, generalized additive model; CART,
classification and regression trees; GBM, gradient boosting machines; BRT, boosted regression machines; GWR, geographically weighted regression; SOM, self-
organized maps; SARIMAX, seasonal autoregressive integrated moving average; MAE, mean absolute error; LiR, linear regression; BI, Breteau index; HI, house index;
CoI, container index; AI, Adult index; KNN, k-nearest neighbors; MSE, mean square error; GT, Google trends; DBSI, dengue Baidu search index; RMSE, root mean
square error; GLM, generalized linear model; ALT, alanine aminotransferase; BTS, bayesian time series; CRF, climate risk factor index; SS, stochastic simulation.
* Corresponding author at: Grupo de Investigaciones Microbiológicas y Biomédicas de Córdoba, Universidad de Córdoba, Montería, Colombia.
E-mail address: whoyos@correo.unicordoba.edu.co (W. Hoyos).

https://doi.org/10.1016/j.artmed.2021.102157
Received 19 November 2020; Received in revised form 8 May 2021; Accepted 17 August 2021
Available online 24 August 2021
0933-3657/© 2021 Elsevier B.V. All rights reserved.
W. Hoyos et al.
disease: non-severe dengue (with or without warning signs) and severe
dengue (SD). This last includes the dengue shock syndrome (DSS) [6].
According to the WHO, more than 350 million dengue virus infections
occur annually worldwide. In addition, 20,000 deaths related to dengue
in the same period of time [7].
Dengue has been the subject of various studies worldwide. Its high
prevalence in tropical and subtropical regions of the world has gener­
ated interest in its diagnosis, treatment and control. Different systematic
literature reviews (SLRs) have been carried out of dengue. Most of them
have been focused on the evaluation of molecules for the generation of
vaccines, control of transmission, epidemiology and development of
rapid-detection tests. In what follows, we briefly explain previous SLRs.
Several SLRs have described the epidemiology of dengue. Jing and
Wang [8] showed the epidemiology of dengue according to its
geographical and temporal distribution. Besides, Jing and Wang evalu­
ated risk factors for transmission and control of dengue. Alhaeli et al. [9]
conducted a review of the epidemiology of dengue in Saudi Arabia,
where environmental conditions are extreme. Other reviews on the
epidemiology of dengue have been carried out in different countries,
such as Pakistan [10], Thailand [11], Malaysia [12], Philippines [13],
Mexico [14] and Brazil [15]. Finally, Villar et al. [1] conducted a SLR of
the epidemiological trends of dengue, in Colombia, for over 12 years
(2000− 2011).
Another group of SLRs has focused on the production of rapid-
detection tests and vaccines against the virus. For instance, Lim et al.
[16] and Luo et al. [17] conducted reviews and meta-analyses to assess
the economic impact of rapid-screening tests. Reviews have also been
conducted to identify the latest economic studies of dengue vaccination
[18,19]. For the development of dengue vaccines, it has been evaluated
the immunogenicity, safety and efficacy of the vaccine [20,21,22].
In recent years, with the emergence of machine learning and the
increase in data generation, computational methods have been devel­
oped for the prediction and evaluation of disease-transmission dy­
namics. This has generated interest for SLRs on this subject, to know the
latest developments and opportunities in this domain. As an example,
Louis et al. [23] developed an SLR of dengue to identify the main
modeling approaches of the disease risk. Another SLR on computational
methods was conducted by Naish et al. [24], focusing on quantitative
modeling with respect to climate change. Andraud et al. [25] conducted
a review of deterministic models of dengue transmission to identify
features for future models. Finally, Lourenço et al. [26] published a
review of the challenges in dengue research from a computational
perspective. The authors focused on real-time data collection, genetic
analysis and integrative modeling approaches. Particularly, integrative-
modeling approaches simulate the epidemiology and molecular evolu­
tion of the virus.
We present a review of three modeling approaches of dengue:
diagnostic, epidemic and intervention. The goal is to present the
development of machine learning models for these contexts. The first
approach is to determine whether a patient has dengue or any of its
variants. The second is to analyze the population-level dengue epidemic;
in addition, to study morbidity and mortality rates. The third is to
analyze the impact of interventions to mitigate epidemics of dengue. To
date, there is no SLR that studies these three aspects related to the dis­
ease together. In addition, it is the first SLR to focus on models to
evaluate the impact of interventions to mitigate dengue epidemics.
Finally, this SLR establishes the state-of-the-art in these approaches,
and, additionally, defines new challenges and opportunities for future
research. The objectives of this SLR are:
• To collect and describe machine learning models for dengue.
• To visualize challenges for future work in dengue modeling.
The present document is structured as follows: Section 2 describes
search and selection process of relevant articles; Section 3 describes
general results of the research; Section 4 discusses the papers, as well as
2
Artificial Intelligence In Medicine 119 (2021) 102157
the challenges and opportunities for research on dengue modeling for
diagnosis, epidemics and interventions to control dengue. The last sec­
tion shows the conclusions, with a description of the works that would
be a priority to develop in this research domain.
2. Methodology
This review was based on the PRISMA methodology [27]. The first
step is to establish research questions; the second is to define a search
strategy to delimit the findings; the third is to select the papers using
eligibility criteria; and, finally, the last is to analyze the articles to
extract strengths, limitations and challenges to overcome. To achieve
the goal of this review, three research questions were proposed:
Q1. Which machine learning models have been developed for dengue
diagnosis?
Q2. Which machine learning models have been developed for the
analysis of dengue epidemics?
Q3. Which machine learning models have been developed for the
evaluation of dengue control strategies?
2.1. Search strategy
We used several digital libraries (databases): ScienceDirect, IEEE
Xplorer, Google Scholar, Emerald, Taylor & Francis and Pubmed. The
inclusion criteria for the selection of publications were: i) articles from
January 2015 to March 2021, in the English language, related to the
development and implementation of diagnostic, epidemic and inter­
vention models of dengue; ii) articles that match the search terms that
describe the research questions. The criteria that allow discarding
publications were: i) articles representing the personal opinions of in­
dividual experts, ii) conference papers, posters, abstracts, short articles
and unpublished works; and iii) articles using ordinary-differential-
equations models and other deterministic approaches. Table 1 shows
the search strings derived from the research questions. Search strings
were structured using the logical operators “OR” and “AND”.
2.2. Selection procedure
The selection procedure was carried out in three stages using inclu­
sion and exclusion criteria above mentioned: i) We chose articles by
evaluating their title and keywords to exclude any non-relevant work.
We also removed duplicate papers; ii) We examined the summaries of
candidate papers from Stage 1. Then, we evaluated each paper to define
if it is selected to the next step; iii) We evaluated the full texts of selected
papers from Stage 2 to exclude papers that did not meet the criteria.
Fig. 1 shows the flowchart of the selection process.
A total of 19.327 papers were recovered from the scientific libraries.
After reviewing the title and keywords, and removing duplicate ele­
ments (Stage 1), 418 papers were selected. Stage 2 consisted of abstract
review, which allowed the selection of 203 papers. Finally, 64 articles
met all the eligibility criteria (Stage 3), where 27 were about diagnostic
modeling, 29 about epidemic models and 8 about prescriptive
Table 1
Search strings used for each research question.
Question Approach String search
Q1 Diagnostic [(diagnosis OR diagnostic OR infection) AND (gender
OR age OR phenotype OR race OR “clinical profile”)
AND model AND dengue]
Q2 Epidemic [(epidemic OR outbreak) AND (predictive OR predicting
OR prediction) AND model AND (dengue OR aedes)]
Q3 Intervention [(fumigation OR vaccine OR “biologic control” OR
“decision making” OR intervention OR prescriptive)
AND model AND (dengue OR aedes)]
W. Hoyos et al. Artificial Intelligence In Medicine 119 (2021) 102157

Fig. 1. Flowchart of the selection process.

(intervention) modeling.
2.3. Preliminary analysis
This subsection shows the preliminary results of selected articles.
Fig. 2 shows the distribution of the reviewed articles on dengue
modeling around the world. The highest number of studies were from
Taiwan with 9 (14%) articles, followed by Brazil and Vietnam, with 7
(11%) articles in both countries. It was expected that Taiwan, Brazil and
Vietnam would be in the top positions. First, because they are endemic
countries where the amount of data available is greater. Second, because
they are countries close to the equatorial axis and contain tropical and
subtropical regions. In contrast, there are dengue-endemic countries
with a low amount of publications on diagnostic, epidemic and

Taiwan
Vietnam

Brazil

0 2 4 6 9
Frequency 1 3 5 7

Fig. 2. Worldwide distribution of the reviewed papers on dengue modeling.

3
W. Hoyos et al. Artificial Intelligence In Medicine 119 (2021) 102157

intervention modeling. Among these are Colombia, Ecuador and developed on diagnostic, epidemic and intervention modeling of
Venezuela. These countries may not have many publications because dengue.
they invest very low in science and research [28].
Many predictors/features/variables are currently used for dengue 3.1. Diagnostic models of dengue
modeling, which are classified according to their own characteristics
and their method of collection. We classified them as: demographic, This section analyzes articles related to diagnostic models of dengue.
economic, clinical, laboratory, environmental, climatic, among others. The analysis was carried out according to different aspects of the disease
For a better understanding, a brief description of them, with some ex­ that are currently important: early detection, seroprevalence used to
amples, is shown in Table 2. In general, the most used variables were determine populations at risk to acquire the disease, use of cytokines and
socio-economic and demographic data. Their easy access and avail­ plasma leakage as early markers of severity, and new diagnostic
ability would explain their high frequency of use. The combination of methods such as Raman spectroscopy.
clinical and laboratory data was the most used for diagnostic modeling.
The use of this data type is crucial for this approach because it allows 3.1.1. Early diagnosis of dengue
finding relationships between the data and early detection of the dis­ Early diagnosis of dengue could prevent complications and death.
ease. In terms of the epidemic approach, the most widely used pre­ For this reason, Macedo-Hair et al. [29] presented the analysis of clinical
dictors/variables were climatic, environmental and meteorological. profiles of 523 dengue patients. In this case, Macedo-Hair et al. used
These data types are widely used for the spatial-temporal analysis of unsupervised learning to find natural clusters associated with clinical
dengue to map the distribution of the mosquito or disease. Finally, patterns in confirmed cases of dengue. The results showed that the
intervention modeling focused on the evaluation of control strategies of model can classify dengue into four states (4 clusters): dengue without
the mosquito. For this reason, entomological data were the most used for warning signs, dengue with warning signs, SD and an intermediate state.
this purpose. These clusters can be used as risk criteria to diagnose dengue. Fernandez
According to this review, the least used variables for dengue et al. [30] presented a model based on logistic regression (LoR) to
modeling are genomics, cellphone and thermal-imaging data. Genomics differentiate dengue from other febrile diseases. The use of laboratory,
data obtained from genetic tests were not usually performed for the clinical and demographic data was useful to reveal the association of
diagnosis of dengue. Cellphone data are not easily acquired due to user predictive variables with the risk of suffering dengue. The results
privacy issues, and thermal imaging requires specialized tools that are showed that there was a strong association of dengue with the explan­
not available in clinical practice. atory variables: male sex, petechiae, skin rashes, myalgias, retro-ocular
pain, positive tourniquet test and gingival bleeding. The accuracy of the
3. Analysis of reviewed papers model to diagnose dengue was 69.2%.
Artificial neural networks (ANN) are machine learning algorithms that
Sixty-four articles were reviewed and analyzed to find what has been describe functional dependencies between input and output variables.
One drawback of ANN is their optimization functions. ANN is set-up as a
Table 2
non-convex optimization problem where there could be a local mini­
Description, examples and frequency of predictors used in dengue modeling in mum that is not a global minimum. To overcome this, Chatterjee et al.
the reviewed papers. [31] used an ANN (multilayer perceptron (MLP)) with a Cukoo search
algorithm (ANN-MCS), to classify healthy people, patients with dengue
Type of predictor Definition Examples Articles
and SD. The results showed that ANN-MCS improves the model accuracy
Demographic + Characteristics related Age, sex, population, 31 (95.65%), compared to the use of unmodified ANN (87.5%). Gambhir
Social + to the development of a housing type, socio-
Economic + population, from a economic level.
et al. [32] also used an ANN (MLP) to early predict dengue. The authors
Population quantitative supplemented ANN with the particle swarm-optimization (PSO) algo­
perspective. rithm. The results of the combined model showed an accuracy of 87%,
Climatic + Characteristics related Temperature, 29 higher than the accuracy of ANN without PSO (79%). A deep neural
Environmental + to climate and rainfall,
network (DNN) is an ANN with several hidden layers between the input
Meteorological + environment. precipitation,
Topographic elevation. and output layers. This kind of ANN can model more complex nonlinear
Laboratory Analytical Platelet and 25 relationships [33]. Ho et al. [34] used a DNN to identify laboratory-
determinations of leukocyte count, confirmed dengue cases using only four input variables (age, body
metabolites in the hematocrit, albumin, temperature, leukocyte count and platelets). The developed model by
blood that may be transaminases
altered in patients with
Ho et al. was compared to LoR and DT. The area under the curve (AUC)
dengue. was used to evaluate the performance. The results showed similar per­
Clinical Signs and symptoms of Blood pressure, fever, 21 formance in the developed models (DNN = 0.86 Vs. DT = 0.85 Vs. LoR
patients with dengue. joint pain, headache, = 0.84), with DNN being slightly better.
retro-ocular pain,
Park et al. [35] developed models to classify dengue, SD and DSS
arthralgia, myalgia.
Search Index and Data from Internet Google Trends, Baidu 7 using structural equation (SE) modeling. Park et al. used clinical and
social networks Search Index, laboratory data for this purpose, and their models showed good per­
Twitter. formance for each disease variety (dengue: AUC = 0.84, SD: AUC =
Entomological Data related to the Breteau index, 7 0.67, DSS: AUC = 0.70). Finally, Khosavanna et al. [36] developed two
biological vector and container index,
its propagation. house index, one
diagnostic algorithms (LoR and DT) based on clinical symptoms of
adult index, dengue patients. The models performed similarly in specificity (LoR =
predations rates. 0.63 Vs. DT = 0.67) and sensitivity (LoR = 0.78 Vs. DT = 0.77).
Genomic Genetic data Gene-expression 2
levels
3.1.2. Seroprevalence of dengue
Thermal Images Images obtained from Thermograms 1
infrared cameras Dengue seroprevalence studies allow knowing past or current cir­
Cellphone Data obtained from Geo-localization 1 culation of the virus in a specific area. This allows, among other things,
cellphones to determine the populations at risk for a disease, and to evaluate the
Mobility Air-passengers travel Destination country 1 mechanisms of transmission [37]. Al-Raddadi et al. [38] developed the
data
first multivariate model using LoR to estimate dengue seroprevalence in

4
W. Hoyos et al.
four endemic cities in Saudi Arabia. The authors analyzed the associa­
tion of risk factors (demographic, clinical, and environmental) with the
disease using a multivariate LoR. They used the odds ratio (OR) with a
95% confidence interval (CI) to present the results. The predictors asso­
ciated with the highest seroprevalence rate were: age over 30 years (OR
[95% CI] = 3.91[2.78, 5.50]), housing type (OR [95% CI] = 1.93[1.62,
2.31]), absence of pest-control activities (OR [95% CI] = 1.39 [1.13,
1.72]) and presence of mosquitoes at home (OR [95% CI] = 1.39 [1.14,
1.70]). Aguas et al. [39] used random forest (RF), on laboratory data
(antibody titration), to estimate the proportion of asymptomatic dengue
in children. The algorithm presented an accuracy of 99.45%, correctly
classifying 361 cases out of 363.
3.1.3. Cytokines
Cytokines are molecules that increase their levels in the blood after a
severe infection. Jayasundara et al. [40] presented a study describing
the role of cytokines in SD: platelet activating factor (PAF), sphingosine 1-
phosphate (S1P), interleukin-1 beta (IL-1β), tumor necrosis factor-alpha
(TNFα) and interleukin-10 (IL-10). They used fuzzy-logic (FL) to di­
agnose SD. The patients were analyzed 96, 108 and 120 h from onset of
fever, using blood levels of cytokines for each time point. The developed
model showed the best accuracy after 108 h from the onset of fever
(85%). Low et al. [41] evaluated vascular endothelial growth factor
(VEGF) and pentraxin-3 to classify the disease into severe and non-
severe. The proposed diagnostic model using LoR showed 76.2% and
73.6% of sensitivity and specificity, respectively. According to the re­
sults, pentraxin-3 is not useful to differentiate SD from non-severe
dengue. There was no significant difference between the two groups
for this cytokine's blood levels.
3.1.4. Raman spectroscopy
In recent years, Raman spectroscopy has been used for medical di­
agnoses, such as cancer [42,43], liver diseases [44,45], and infectious
diseases, such as tuberculosis [46] and Chagas disease [47]. Khan et al.
[48] proposed the extraction of the Raman spectrum, in serum samples,
from healthy people and dengue patients. The goal was to classify the
samples into normal and pathological using RF. The model had a good
performance (accuracy = 91%). The same authors published another
paper [49], but this time, they used a support vector machine (SVM) for
the classification task. Different kernels were used: linear, polynomial
and radial. The best kernel was the grade one polynomial; however, the
performance with SVM was lower (85%) compared to RF (91%).
3.1.5. Plasma leakage and severity
Plasma extravasation is a warning sign for SD and should be detected
early to avoid complications and death. This sign is characterized by
serous effusions at the level of various cavities, such as pleura, peri­
cardium and peritoneum. For this reason, Suwarto et al. [50] developed
a scoring system to detect pleural and/or ascitic effusion. Suwarto et al.
implemented a LoR using laboratory data. To each factor, a score is
assigned to determine the risk of plasma leakage. The higher score is
assigned to a patient, the more likely the patient is to leak plasma. The
developed model detected plasma leakage or ascites, with an accuracy of
77.4%. Another study by da Silva et al. [51], used the same regression
technique (LoR) to evaluate risk factors for hospitalization after dengue
infection. The explanatory variables used were demographic, clinical
and laboratory. The authors found that multi-organ failures are the most
influential factors in hospitalization (OR[95% CI] = 5.75[3.53, 9.37]).
Fernandez et al. [52] used a multivariate logistic model, in Honduran
patients, using demographic, clinical and laboratory data. Fernandez
et al. used plasma leakage as the target variable since this is the main
warning sign of SD. The developed model achieved an accuracy of
70.9%, with a sensitivity of 76.4% and a specificity of 70.3%. The same
modeling technique was used by Phuong et al. [53], with the addition of
free plasma deoxyribonucleic acid (DNA) as a predictor variable. The
model achieved a sensitivity of 87.5%, and a specificity of 54.7%.
5
Artificial Intelligence In Medicine 119 (2021) 102157
Davi et al. [54] used gene expression data to diagnose SD. The au­
thors used an MLP model with an average accuracy of 86%. Another
study conducted by Tuan et al. [55], in Vietnam, applied multivariate
LoR models to demographic, clinical and laboratory data. The AUC re­
sults were 0.95, with a sensitivity of 87% and specificity of 88%.
Another model, using the same data and technique, was developed by
Ahmad et al. [56]. The authors evaluated the main warning signs for SD.
The best results showed that the model had a sensitivity of 91% when at
least one warning sign was present, and a specificity of 99% when there
were more than 5 warning signs. The study of Phakhounthong et al. [57]
presented the use of decision trees (DT) for SD in children. The model was
based on clinical and socio-demographic data of dengue patients. The
sensitivity, specificity, and accuracy of the model were 60.5%, 65% and
64.1%, respectively. Finally, Huang et al. [58] developed several models
to diagnose severe dengue using demographic information and
laboratory-test results. Huang et al. applied several machine learning
techniques, such as LoR, RF, GBM, SVM and ANN. The best model was
ANN with an accuracy of 75% and an AUC of 0.83.
Zhang et al. [59], in their study, showed a new variable to diagnose
SD: the aspartate aminotransferase/platelet count ratio index (APRI). The
authors developed an LoR to evaluate the performance of APRI, in
conjunction with other laboratory variables, such as prothrombin time
and leukocyte count. The model performed well, reporting an AUC of
0.87. Another work carried out by Lin et al., [60] also used LoR, but
using hyaluronic acid as a feature or variable. The developed model by
these authors had a moderate performance, with an AUC of 0.69,
specificity of 55%, and sensitivity of 76%. Lee et al. [61] implemented
LoR models for the development of a clinical-risk score for early diag­
nosis of SD. The researchers proposed that the coefficients in the model
can be used as a risk score. The best developed model obtained an AUC
of 0.92. The sensitivity and specificity of the model were 80.3% and
85.8%, respectively.
DSS is a potentially life-threatening complication of the disease. Lam
et al. [62] proposed a diagnostic model to detect DSS in children. The
authors developed LoR models to determine the relationships between
clinical and laboratory variables with the presence of DSS. In addition,
alternative techniques, such as adaptive least absolute shrinkage and se­
lection operator (LASSO), generalized additive model (GAM), classification
and regression trees (CART) and gradient boosting machine (GBM), were
compared. The logistic model performed favorably (AUC = 0.74),
compared to the alternative modeling strategies (LASSO = 0.73, CART
= 0.61, GAM = 0.69, GBM = 0.72). Another research work using LoR
was carried out by Lam et al. [63], in Vietnam, where they evaluated the
platelet count for the diagnosis of DSS. The model was better with the
use of platelet count than without it (AUC = 0.73 Vs. AUC = 0.66).
In summary, most of the modeling approaches for dengue diagnosis
were based on LoR (see Fig. 3). Logistic models are widely used in the
health field because of their simplicity to perform and interpret the re­
sults. These models were developed primarily to assess the factors
associated with the risk of dengue infection, and to determine the as­
sociation of predictive factors with disease severity. The categories of
variables most frequently used to construct these models were de­
mographic, clinical and laboratory parameters (see Fig. 3). This result is
possibly due to the fact that these types of features are the most available
in all countries with mandatory surveillance systems. GBM, FL and SE
models are possibly the least used because of their mathematical
complexity and difficulty to implement. Finally, the reviewed models for
diagnosis were implemented as classification tasks: They only deter­
mined whether the patient had the disease or not. Diagnostic models
should go further, and evaluate causal relationships among predictors
and dengue.
3.2. Epidemic models of dengue
This section analyzes epidemic modeling approaches of dengue. This
section was divided into subsections according to key aspects, such as
W. Hoyos et al.
40
30
Frequency
20
10
LoR ANN RF DT
Model
Fig. 3. Types of models vs. type of data sources for diagnostic models. The frequency indicates the number of times the technique was implemented in the studies.
Abbreviations: LoR = Logistic Regression, ANN = Artificial Neural Networks, RF = Random Forest, DT = Decision Trees, SVM = Support Vector Machines, GBM =
Gradient Boosting Machines, FL = Fuzzy Logic, SE = Structural Equations.
the type of analysis performed (e.g., spatial-temporal analysis), the types
of data used (e.g., the Internet data from the social networks and search
indexes). Finally, there is an important section dealing with the pre­
diction of mortality, a latent problem that should be addressed with
predictive modeling.
3.2.1. Spatial-temporal analysis of dengue
The spatial-temporal analysis of dengue is the most studied field of
the disease, among the reviewed articles. There are many studies that
use machine learning to evaluate the spread of vectors and diseases.
Rossi et al. [64] used boosted regression-trees (BRT) to conduct a spatial-
temporal analysis of dengue with data from 76 countries. BRT is a
modeling technique used primarily in ecology to explain or predict a
phenomenon. The data collected were temperature, rainfall, migration
and population density. The study showed that higher population den­
sity and shorter distances between countries with dengue outbreaks are
relevant factors that characterize the disease. Geographically weighted
regression (GWR) was used by Delmelle et al. [65], to evaluate the role of
environmental and socioeconomic determinants of dengue in Cali,
Colombia. The authors found that socioeconomic status, population
density, proximity to both tire shops and plant nurseries; and the pres­
ence of sewage systems, are related to the disease. Mao et al. [66] used
RF to predict the presence of dengue cases in a given area using topo­
graphic, climatic and population data. Since people are more likely to
become ill when they travel to other locations, an important contribu­
tion of this work is the use of cellphone tracking data. Based on this, the
authors reported an accuracy of 95%. Finally, Mutheneni et al. [67]
mapped the levels of dengue endemicity in some districts of India, to
identify groups at risk. The authors used self-organizing maps (SOM),
with environmental data, for this purpose. The results indicated that the
districts of Warangal, Karimnagar, Khammam and Vizianagaram are hot
spot regions.
6
Artificial Intelligence In Medicine 119 (2021) 102157
Data source
Clinical
Demographic
Environmental
Genomic
Laboratory
Raman spectrograms
Social
SVM GBM FL SE
A study by Akter et al. [68] used linear regression (LiR), with
ecological and socio-demographic factors, to observe the spatial-
temporal trend of dengue in Australia. The results of regression anal­
ysis showed an increased trend of dengue incidence with some factors,
such as housing types and households with rainwater tanks. Yue et al.
[69] and Reyes-Castro et al. [70] used the same technique in five dis­
tricts of China and two arid cities of Mexico, respectively. The two
studies used environmental and spatial data to build the models. Be­
sides, socioeconomic data were aggregated for model improvement. On
one hand, Yue et al. indicated the factors and dengue outbreak were
significantly positively correlated. On the other hand, Reyes-Castro et al.
showed that transmission foci started in neighborhoods with high-
population density and low access to health services.
In Brazil, according to the Ministry of Public Health, a year is
epidemic, in a city, if the incidence is greater than 100 cases per 100,000
inhabitants [71]. In this regard, Stolerman et al. [72] developed an SVM
to predict whether a year will be epidemic or not. The data used by
Stolerman et al. were climatic and epidemiological of 16 years. The SVM
predicted the epidemicity of a year at 91% accuracy.
Several studies have compared the performance of different machine
learning models to predict dengue burden, outbreaks and importation of
dengue into Europe. Carvajal et al. [73] compared machine learning
models, using weather factors, in the Philippines. The objective was to
find out which meteorological factors were the best predictors of dengue
in that country. The techniques used were GAM, seasonal autoregressive
integrated moving average with exogenous variables (SARIMAX), RF and
GBM. They reported that relative humidity is the most important
meteorological factor in the model. The highest performance, in terms of
mean absolute error (MAE), was RF (0.23), followed by GBM (0.24). Zhao
et al. [74] compared RF and ANN to predict dengue burden, in
Colombia, at national and local scales. The comparison between the
models was performed using MAE, and the results showed that RF (0.86)
W. Hoyos et al.
performs slightly better than ANN (0.95). According to the level of
prediction, the results showed that RF performs better at the national
level than the sub-national level, demonstrated by lower MAE values in
12-week forecasts (national = 0.86 Vs. local = 0.97). Salim et al. [75]
used various machine learning techniques (SVM, DT, ANN, Bayes
Network) to predict epidemics in Malaysia. Climatic variables were used
as a predictor to build the models. According to the results reported by
Salim et al., linear SVM performed the best, with an accuracy of 70%,
specificity of 95% and sensitivity of 14%, when using the original data.
On the other hand, class balancing improved the sensitivity to 64%.
Finally, Salami et al. [76] developed and compared machine learning
models to predict dengue importation into Europe. Salami et al. used air-
passenger data to create connectivity indices between a source and
destination country. The techniques implemented were RF, GLM, GBM
and partial least squares. GBM had the best performance, with an AUC of
0.94, a sensitivity of 0.94 and a specificity of 0.93.
3.2.2. Distribution of the vector
Simulation of vector distribution that transmits dengue is important
to establish control strategies by health authorities. For this type of
modeling, entomological data, such as the Breteau index (BI), house index
(HI), container index (CoI) and adult index (AI), are commonly used. BI is
the number of positive containers per 100 houses inspected. HI is the
percentage of houses infested with mosquito larvae or pupae. CoI is the
percentage of water containers infested with mosquito larvae or pupae.
Finally, AI is the number of female mosquitoes captured divided by the
number of houses inspected [77,78].
Parra et al. [79] developed a GAM using BI and mosquito genetic
data. Meteorological data were also used to build the model. The pro­
posed model required 71.5% fewer human and operational resources
than the BI measurement. Similarly, Chang et al. [77] used entomo­
logical indices as a tool for early prediction of a dengue epidemic. The
implemented regression models obtained accuracies of 83.8, 87.8, 88.3
and 88.4%, for BI, HI, CoI and AI, respectively. Ding et al. [80] simulated
the distribution of A. aegypti and A. albopictus using environmental,
climatic, social data and three machine learning methods (SVM, GBM
and RF). Models with RF performed better followed by GBM; however,
there were no significant differences between the results of AUC
(A. aegypti: 0.973 Vs. 0.974 and A. albopictus: 0.971 Vs. 0.972). Jacome
et al. [81] used LiR to identify the most important risk factors for the
distribution of A. aegypti in a coastal zone in Ecuador. Environmental
and spatial data were used for this purpose. Temperature and population
density were the factors most likely to predict the number of cases.
Modeling of mosquito breeding sites, using remote sensing, is gain­
ing interest in the scientific community. Scavuzzo et al. [78] imple­
mented several machine learning techniques to model the oviposition
activity of A. aegypti, using time series obtained from satellite image
data. The use of these techniques allowed finding non-linear relation­
ships between environmental variables and the oviposition of A. aegypti.
The techniques used were: SVM, MLP, k-nearest neighbors (KNN) and DT.
The evaluation of the models was done with mean square error (MSE).
The results showed that the best model was KNN (MSE = 0.49), followed
by MLP (MSE = 0.52), SVM (MSE = 0.61) and DT (MSE = 0.77).
3.2.3. Search-index data
The use of Internet searches has become a useful tool to predict
disease outbreaks, where Google trends (GT) is the reference in this field.
GT is a tool from Google that displays the most popular search terms in a
fixed time and location. Data associated with these searches is used for
prediction.
The research conducted by Wu et al. [82] used climatic data from
Taiwan combined with GT data. The model was built with DT and the
findings revealed that temperature and humidity were the most relevant
factors, with the greatest power of classification, while age and gender
were the least relevant. Wu et al. [82] found that the use of GT data
decreases the accuracy of the model (96% Vs. 94%). Strauss et al. [83]
7
Artificial Intelligence In Medicine 119 (2021) 102157
compared the accuracy of GT with conventional surveillance systems, in
Venezuela, for 10 years. The authors used LiR to predict the cases, re­
ported officially by the Ministry of Health, based on GT data. The overall
coefficient of determination (R2) was 0.75.
In countries such as China, GT is not available to users, but there are
alternatives, such as Baidu Search, a search engine that stores the
searches made by users. Li et al. [84] used the data from Baidu index
database and calculated the Dengue Baidu search index (DBSI) to improve
the prediction of local dengue epidemics in Guangzhou. Climatic data
also were used by Li et al. to train a GAM. The model performance was
evaluated with the root mean square error (RMSE). The results showed
that the model with DBSI was better than without DBSI (RMSE = 59.9
Vs. RMSE = 203.3). Another study, conducted in China, by Liu et al.
[85], used regression trees on DBSI data. The results demonstrated a
strong association between DBSI and dengue incidences. The accuracy of
the models was above 90%.
3.2.4. Social networks
Social networks provide information on the mobility of individuals in
a population because a large percentage of social-network data is geo-
tagged. According to this review, Twitter is the most used social
network to predict dengue. Marques-Toledo et al. [86] used tweets to
predict dengue, at local and national levels, in Brazil. Social network
data were supplemented with demographic and incidence variables. The
model had the ability to predict an outbreak up to 8 weeks in advance,
with an MAE of 0.35. Another similar work was carried out by Ram­
adona et al. [87], who used geo-tagged data from Twitter and a gener­
alized linear model (GLM). The main objective of this research was to
predict the risk of dengue in Yogyakarta, Indonesia. The model yielded
an RMSE value of 0.78 when including Twitter data with the dynamic
index of incidence weighted by mobility. Finally, Souza et al. [88] used
Twitter data to create unsupervised models that detect spatial clusters to
characterize high-risk regions of dengue.
3.2.5. Prediction of morbidity and mortality
The mortality rate in patients with SD is too high, mainly in children
and geriatric patients [89,90]. For this reason, it has been important to
develop models to predict morbidity and mortality. Kesorn et al. [91]
applied various machine learning techniques to predict the morbidity
rate of SD. DT, KNN, SVM (with linear, polynomial and radial kernels)
and ANN were applied to climatic and demographic data. A. aegypti
infection rates were also used to improve model performance. The re­
sults showed that SVM with radial kernel had better accuracy, with
88.4%, when the infection rate in the mosquito was added. Md-Sani
et al. [92] developed an LoR model in Malaysia. The goal was to iden­
tify risk factors that would allow the prediction of mortality. The best
developed model was with age, serum bicarbonate, serum lactate and
alanine aminotransferase (ALT), with an AUC of 0.84.
Huang et al. [89] used demographic, clinical and laboratory data of
patients over 65 years of age (N = 627). They used an LoR model to
estimate the mortality of dengue. The model predicted a mortality of
57.1% when at least two predictors were present in dengue patients.
Huang et al. [90] conducted another study with a larger sample (N =
2358). Huang et al. developed a similar model to the one developed in
[89] to assign a score to each patient, to know its probability of death.
The results showed that the model had an AUC of 0.85.
3.2.6. Thermal images
Nagori et al. [93] used thermal imaging for the prediction of DSS.
The authors used images of pediatric patients to train a GLM. The
developed model demonstrated the usefulness of thermal imaging to
predict DSS, with an AUC of 0.76.
In summary, compared to diagnostic models, where LoR was most
commonly used (see Fig. 3), in epidemic models, there is a variability in
the frequency of the modeling approaches used. Although LiR was the
most used approach for prediction, other techniques, with a high
W. Hoyos et al. Artificial Intelligence In Medicine 119 (2021) 102157

frequency, such as RF, SVM, LoR and GAM/GLM, were also imple­
mented. Spatial-temporal analyses were used for the case of prediction
using regression approaches. The most used technique for this task was
LiR, where several studies used 16 times this type of modeling. As we see
in Fig. 4, there is a relationship between the use of data types and
modeling types. The LiR models used socioeconomic, demographic and
environmental data. This type of data has been widely used for LiR
because it allows mapping the distribution of the mosquito and the
disease. These types of data were used in almost all the epidemic
modeling approaches (see Fig. 4), except with LoR, which did not use
environmental data. Clinical and laboratory data were mainly used with
LoR, a technique widely used by medical personnel and epidemiologists
to predict the presence or absence of a disease (classification task, 9
papers). ANN have been little used, probably because of the low-quality
and availability of data for dengue. It has been demonstrated that the
performance of ANN is directly proportional to the quantity and quality
of data [94]. Cellphone data have been little used to map dengue risk.
The limited availability of data and the problem of data privacy could be
the reason for this inconvenience. Finally, the use of GAM/GLM has
increased with data extracted from the Internet (search indexes and
social networks). These non-parametric models are increasing their use
because they can capture features of a non-linear nature from unstruc­
tured data, such as trend data or tweets.
3.3. Strategy evaluation models to control dengue
Few intervention models have been developed for the evaluation of
dengue control strategies. In this subsection, we analyze the articles
according to different approaches, such as biological control using co­
pepods, entomopathogenic fungi, Wolbachia strains, vaccination and
fumigation.
3.3.1. Copepods
In recent years, biological control of the vector A. aegypti and
A. albopictus has emerged using biological predators called copepods.
These are crustaceans with the ability to systematically devour young
mosquito larvae [95]. Kalimuthu et al. [95] used a GLM to evaluate the
predation efficiency of Mesocyclops formosanus on young larval pop­
ulations. The developed model showed the effectiveness of using co­
pepods to control the vector, and thus the disease. Another study, by
Udayanga et al. [96], also used a GLM to compare the predation effi­
ciency rates of five copepods on Aedes larvae. According to the model,
the highest predation efficiency rates were higher with Mesocyclops
leuckarti, with 17.45% and 16.75% for A. aegypti and A. albopictus,
respectively.
3.3.2. Entomopathogenic fungi
Entomopathogenic fungi produces diseases and cause the death of
insects and arthropods. These types of fungi are a useful alternative to
control the mosquito that transmits dengue. Lee et al. [97] built a GLM
to evaluate the pathogenic activity of six species of fungi (Beauveria,
Cordyceps, Metarhizium, Paecilomyces, Purpureocillium, and Verticillium)
on A. albopictus. The model showed that Metarhizium anisopliae had the
highest activity to eliminate A. albopictus, with a mortality rate of 73%
after two days, and 90% after 5 days.
3.3.3. Wolbachia strains
Wolbachia is a bacterium that naturally infects insects. Infection of
males with this bacterium produces a generation with unviable offspring
when mated with an uninfected female [98]. This approach has been

15
Data source
Cell Phone
Clim. + Env. + Met. + Top.
Clinical + Lab.
Entomological
GT + DBSI + Twitter
Mobility
Soc. + Econ. + Dem. + Pop.
10 Thermal Images
Frequency

LiR RF SVM LoR GAM/GLM DT GBM ANN GWR KNN BRT Other
Model

Fig. 4. Types of models vs. type of data sources for epidemic models. The frequency indicates the number of times the technique was implemented by the studies.
Abbreviations: Clim = Climatic, Env = Environmental, Met: Metrological, Top = Topographic, Lab = Laboratory, GT = Google Trends, DBSI = Dengue Baidu Search
Index, Soc = Social, Econ = Economic, Dem = Demographic, Pop = Population, LiR = Linear Regression, RF = Random Forest, SVM = Support Vector Machines, LoR
= Logistic Regression, GAM = Generalized Additive Model, GLM = Generalized Linear Model, DT = Decision Trees, GBM = Gradient Boosting Machine, ANN =
Artificial Neural Networks, GWR = Geographically Weighted Regression, KNN = K-Nearest Neighbors, BRT = Boosted Regression Trees.

8
W. Hoyos et al. Artificial Intelligence In Medicine 119 (2021) 102157

developed as a method to control the spread of Aedes. Nazni et al. [99] modeling. GLM was implemented with entomological, climatic and
used a Bayesian model (Bayesian time series (BTS)) to estimate the population data, where entomological data were the most used. A sto­
reduction of dengue cases in Malaysia after infecting mosquitoes with chastic simulation was only used in one work, with entomological and
Wolbachia wAlbB strain. The model estimated dengue case reduction of population data. Finally, BTS was implemented in one work with
40.3% in intervention sites. Other authors have also developed models entomological data. Entomological data is very frequent because dengue
to evaluate the impact of Wolbachia against Aedes; for instance, Indriani control is based mainly on the control of the mosquito, the vector of
et al. [100] used a GLM model to estimate the effect of Wolbachia (wMel dengue transmission. Fig. 5 shows the types of data used for each
strain) over the reduction of dengue incidence in Indonesia. The intervention modeling approach.
developed model was able to reduce dengue incidence by 73% (95% CI:
49%–86%). Finally, Ryan et al. [101] used the same model and the same
3.4. A general analysis of data types and machine-learning techniques for
Wolbachia strain, and achieved a 96% (95% CI: 84%–99%) reduction of
dengue modeling
dengue incidence in Australia.
The types of data and techniques considered in the reviewed studies
3.3.4. Vaccination
were analyzed by question (see previous sub-sections); however, it is
Another option for dengue control is vaccination. Lee et al. [102]
important to analyze them globally. Fig. 6 shows the intersection be­
implemented a GLM for validation between the climatic risk factor index
tween the types of data and techniques used in all the reviewed studies.
(CRF) and dengue incidence, to estimate the vaccination coverage rate
In this figure, we may see that the predominant technique used was LoR,
and the number of doses required. CFR index was created using 12-
and, as we comment in the section of diagnostic models, this technique
month moving averages of climatic and non-climatic factors. The cli­
was frequently used with clinical data and laboratory results. LoR
matic factors were temperature, precipitation and humidity. The non-
models are often used by medical personnel to diagnose dengue. Ac­
climatic factors were population, density and elevation. The study was
cording to Fig. 6, the second is RF, with different data sources used by
conducted in Colombia, Thailand and Vietnam, and the estimated
this technique, such as climatic, clinical and sociodemographic vari­
vaccination coverage rates were 63%, 90% and 91%, respectively.
ables. This variability in the use of predictors/variables is explained
because RF is a technique that was found in diagnostic models and
3.3.5. Fumigation
epidemic models, and can be used for both regression (spatio-temporal
Fumigation has been widely used worldwide to reduce the burden of
analysis) and classification (dengue diagnosis) tasks. Also, Fig. 6 shows
dengue virus-infected mosquitoes. Thus, Hladish et al. [103] con­
that LiR is used in spatio-temporal analysis where logically the utiliza­
structed a stochastic simulation (SS) model to predict the effectiveness of
tion of climatic and sociodemographic variables to map the distribution
spraying in Yucatán, Mexico. The results of the model indicate that the
of the disease or vector was predominant. Finally, another of the most
proactive application of this control method could reduce symptomatic
used techniques was GAM/GLM, a technique mainly used in epidemic
infections by up to 89.7% in the first year, and 78.2% in the five cu­
models and intervention models. For the first case, the objective was to
mulative years.
determine relationships between predictors and dengue incidence, and
In summary, GLM was the most applied technique for intervention
for the second case, the objective was to evaluate the impact of dengue

Data source
Climatic
Entomological
Population

4
Frequency

GLM SS BTS
Model

Fig. 5. Types of models vs. type of data sources for intervention models. The frequency indicates the number of times the technique was implemented by the studies.
Abbreviations: GLM = Generalized Linear Model, SS = Stochastic Simulation, BTS = Bayesian Time Series.

9
W. Hoyos et al.
Thermal Images
Entomological
GT + DBSI + Twitter
Clim. + Env. + Met. + Top.
Genomic
Soc. + Econ. + Dem. + Pop.
Mobility
Clinical + Lab.
Raman spectrograms
Cell Phone
Fig. 6. Types of models vs. type of data sources for all models. The thickness of the links indicates the number of times the technique and data type was implemented
by the studies. Abbreviations: GT = Google Trends, DBSI = Dengue Baidu Search, Clim = Climatic, Env = Environmental, Met: Metrological, Top = Topographic, Soc
= Social, Econ = Economic, Dem = Demographic, Pop = Population, LoR = Logistic Regression, RF = Random Forest, LiR = Linear Regression, GAM = Generalized
Additive Model, GLM = Generalized Linear Model, DT = Decision Trees, SVM = Support Vector Machines, ANN = Artificial Neural Networks, GBM = Gradient
Boosting Machine, KNN = K-Nearest Neighbors, GWR = Geographically Weighted Regression, BRT = Boosted Regression Trees.
control strategies.
4. Discussion
Dengue modeling is a key tool for early detection of dengue, evalu­
ation of risk factors for SD, and may also be useful to control vectors that
transmit the disease. Although extensive works have been done on these
issues, it is important to know what aspects of dengue modeling have not
been worked on, to develop future works that will allow a significant
decrease in disease morbidity rates. The main objective of this work was
to give an overview of diagnostic, epidemic and intervention modeling,
in addition, to determine important challenges for future works.
4.1. Limitations of the studies and challenges
This section is focused on the limitations of the reviewed studies.
Based on those limitations, we describe some research challenges or
opportunities for each of the approaches presented: diagnostic, epidemic
and intervention.
4.1.1. Diagnostic models
The diagnostic models reported in the reviewed articles focused on
the detection of dengue or the differentiation between other diseases,
such as zika, chikungunya and malaria [29,30,35,39,51,59]. This is
useful because the characteristics presented in these diseases, as signs
and symptoms, are also present in dengue. However, it is necessary to go
further and develop cause-effect models that allow a deeper under­
standing of the main causes that lead to high morbidity and mortality
rates of dengue. It is fundamental to develop cause-effect models of
dengue to know the importance of the factors that contribute to the
disease. Specifically, there is a need to understand both the
10
Artificial Intelligence In Medicine 119 (2021) 102157
GAM/GLM
BRT
KNN
LiR
Other
GWR
ANN
DT
SVM
GBM
RF
LoR
interrelationship among the characteristics present in the disease, and
the influence on the variants of dengue. For example, the comorbidities
are basic diseases that can occur jointly with dengue, being the most
common chronic-renal disease and chronic-hepatic disease. In these
diseases, the blood levels of some parameters are elevated, which also
increases in dengue. This aspect must be taken into account to develop
future models of dengue.
According to the reviewed articles, there are many predictors/fea­
tures used for the diagnostic, epidemic and intervention models of
dengue (see Table 2). However, other predictors could be used and
evaluated for this purpose. Raman spectroscopy data would be useful
because the technique has the ability to diagnose the disease early [48].
Another type of predictor that could be used for modeling would be
genomic data. According to this review, only two papers [31,54] have
used this type of data to look for relationships with the disease. Tech­
niques to measure genetic data are expensive and are not usually per­
formed in clinical practice, making it difficult to obtain such data. The
wide collection of genetic data would allow a better understanding of
the dynamics of dengue at the population level, providing key insights
into genetic factors that are difficult to track with clinical records alone.
The use of a large number of variables could be useful to model
dengue; however, a disadvantage of this approach is that the use of too
many descriptors could cause the problem of the curse of dimension­
ality. This is characterized by the high dimensionality of the feature
space where patterns cannot be easily recognized [104]. In addition, this
phenomenon can sometimes hinder the optimization and speed of
execution of the models. To solve this problem, different preprocessing
techniques have been developed. However, in the reviewed papers, very
little use of these techniques was reported (see Fig. 8). The most-
reported preprocessing technique was normalization with 9.4%, fol­
lowed by PCA with 6.3%. This shows that it is necessary to increase the
W. Hoyos et al.
use of preprocessing techniques (feature engineering processes
[105,106]) that allow, among other things, analysis between all the
predictors available, to find the most influential features on the disease,
weighting their influence.
4.1.2. Epidemic models
Epidemic models of dengue are the most published modeling
approach at present, according to this literature review. The largest
percentage of reviewed studies use environmental and climatic data to
analyze the distribution of the mosquito and the disease.
In recent years, there has been an increasing interest to map the
distribution of mosquitoes (A. aegypti and A. albopictus). Knowledge of
the distribution of these vectors can help prevent the disease [99]. For
this task, geographic information systems (GIS) have been used to analyze
the relationship between climatic conditions and the distribution of the
vector [64,79,97,102]. However, this type of information, collected and
used for the models generation, has not a high resolution [80]. For this
reason, a deep-learning approach and high-resolution Google images
could be considered, using significant features (shrubs, urban areas,
roads and puddles) identified from the images. The goal is to predict
regions suitable for mosquitoes on a finer scale.
Data quality is an important aspect for some machine learning al­
gorithms because their performances depend on this feature [107]. The
management of dengue data quality is one of the most important chal­
lenges at present. Most of the databases, with clinical and epidemio­
logical data, reported by surveillance systems, have some problems,
such as incorrect data or missing data [51]. The high demand for
healthcare, in some places, may cause that medical personnel do not
correctly fill out the epidemiological forms provided for this purpose.
This was one of the most common limitations found in the reviewed
studies [35,38,40,53,57,59,93]. When this type of data needs to be
analyzed, this inconvenience sometimes forces the elimination of com­
plete records, which reduces the size of the database. According to the
reviewed articles, none of the papers used models to manage uncertainty
related to data quality. To solve this problem in dengue, several machine
learning alternatives could be used, such as Bayesian models and fuzzy
approaches, which have been used in other domains [108,109,110,111].
Besides, approaches that use robust estimators to deal with the problem
of missing values and outliers have been developed [112]. Another
option could be to generate complementary data to the existing ones.
This consists of creating new data that have similar characteristics (e.g.,
distribution) to the available data. In recent years, also, the generation
of synthetic data has allowed the construction of more robust models in
other areas of knowledge, such as environmental sciences [113,114].
This could also be explored in this domain.
4.1.3. Intervention models
Of the three approaches described in this article, this is the most
promising for future works, because the use of these models, in decision-
making for dengue, is very limited [102]. Below, we show some fields
that have not been much explored on dengue modeling.
Prescriptive models have uncertainty, and being probabilistic, can
lead to incorrect decision-making [115]. Particularly, the uncertainty of
prescriptive models is one of the main challenges in dengue. The key
components of this aspect are the lack of certainty of the model, the
uncertainty in the quality of data, and the subjectivity of the human
being to build the prescriptive model [116]. There is a need to generate
prescriptive models based on data and less based on expert knowledge of
the domain. Additionally, real-time data processing has not been
explored. The developed models are static, and the priority is to use
strategies that have the capacity to process time-varying data. Auto­
mation of prescriptive processes is needed, where the model is always
adjusted in case any inconvenience occurs. In this way, higher-quality
decisions can be made in the shortest possible time to overcome the
problem presented.
Standard machine-learning approaches require the centralization of
11
Artificial Intelligence In Medicine 119 (2021) 102157
training data in one device or data center. This could be a problem if the
data are in different locations. According to this review, all articles used
the centralized machine-learning approach. Most dengue information
systems, in many countries, do not collect their data in one place. To
overcome this drawback, one of the most secure and robust cloud in­
frastructures to process this data, developed by Google, could be used.
The approach, called federated learning, is a kind of collaborative
machine-learning without centralized training data and works as follows
(see Fig. 7): the device downloads the current model, improves it by
learning from the phone data, and then summarizes the changes as a
small focused update. Only this updated model is sent to the cloud, using
encrypted communication, where it is immediately averaged with the
global model to improve the shared model. All training data remains on
the device, and no individual update is stored in the cloud [117].
Finally, no studies were found that have considered implementing
combined models for diagnosis, prediction and prescription. An
important challenge of dengue modeling is the development of these
types of hybrid models. Diagnostic, epidemic and intervention models
combined could be superior in performance to the three models sepa­
rately developed. Additionally, real-time updating for diagnostic and
epidemic models, with the automation of prescriptive model decision-
making, would considerably decrease the uncertainty present in these
types of problems.
4.2. General summary
This section shows a brief summary of the challenges found that
could be used to develop future works. Fig. 8 represents and summarizes
the aspects evaluated in this article. In the upper part, we can note the
aspects evaluated: i) the most used preprocessing techniques, ii) the
applied machine learning tasks, iii) the type of modeling approach, iv)
the technique used and, v) the varieties of dengue that were studied.
Fig. 8 shows the characteristics most and least used to model dengue.
The width of the nodes and links is proportional to the number of
reviewed articles that fall into each of the categories. Different colors
were used in the links, to facilitate the visualization of the connections.
Fig. 8 clearly shows the little use of preprocessing techniques (feature
engineering) in the reviewed articles (45 of 64 articles do not report
preprocessing techniques). Perhaps, the authors assume preprocessing
as an implicit stage within modeling. Also, we can observe that almost
all articles on diagnostic modeling used the classification task to detect
the presence of the disease. The regression task was closely related to
epidemic and intervention models. As mentioned above, intervention
modeling was the least frequent approach in the reviewed articles.
Finally, it can be observed that SSD is the least worked variation of the
disease, mainly because of the low frequency and low quality of the data
related to this syndrome.
For the diagnostic modeling approach, three challenges could be
identified. As we could see in this work, models to detect the disease and
differentiate it from similar ones are quite implemented. Now, it is
fundamental to develop cause-effect models of dengue because they are
necessary to know the importance of the factors that can contribute to
the disease. As can be seen in Fig. 3, several types of features have been
used for the diagnosis of dengue. Adding new predictors for diagnosis
could facilitate and improve detection time for treatment and avoidance
of death. Although the use of many predictors for diagnosis could be
beneficial, it could also cause difficulties with respect to the dimension
of the data. One opportunity is the creation or improvement of pre­
processing techniques that will enable the identification of key charac­
teristics that are most related to dengue.
Although dengue epidemic modeling is the most reported in the
literature, data available for dengue has quality problems. The uncer­
tainty of models based on this type of data is high. One of the main
challenges of epidemic approaches is the development of models that
address and clearly express the associated uncertainty and measure the
reliability of the predictions. It is crucial that improvements are
W. Hoyos et al. Artificial Intelligence In Medicine 119 (2021) 102157

Fig. 7. Description of federated learning. The device personalizes the model locally, based on the usage (A). Many users' updates are aggregated (B) to form a
consensus change (C) to the shared model, after which the procedure is repeated. Source: https://tinyurl.com/y9ykdbve

Fig. 8. Trends of reviewed papers for dengue modeling. Abbreviations: ML = Machine Learning, NR = Not reported, PCA = Principal Component Analysis, SD:
Severe Dengue, DSS = Dengue Shock Syndrome.

developed, or new techniques are created, to generate more accurate Decision-making for treatment and control of the disease depends on
results. this modeling approach. Prescriptive analysis, in general, presents
The frequency of articles on intervention modeling of dengue is low. challenges that must be taken into account. One of these is the low

12
W. Hoyos et al.
number of automatic prescriptive models for data-based decision-mak­
ing systems. The subjectivity of the domain expert could affect the
quality of the decision made, and static models can not handle the
changes that occur. Data-based and up-to-date prescriptions would be a
valuable tool for the treatment and/or control of dengue.
Finally, a challenge common to all three approaches reviewed in this
article is the development of combined models (diagnostic, epidemic
and intervention), to automate the prescription. Autonomous cycles of
data analysis tasks (see [118,119] for more details about this concept),
which integrate the previous models, can assist in decision-making as
quickly as possible. For dengue, this is crucial, due to the high morbidity
and mortality of the disease.
5. Conclusions
We conducted an SLR on dengue modeling based on machine
learning. The main objective was to know about diagnostic, epidemic
and intervention models that have been developed for the disease. Sixty-
four articles were selected and analyzed from several scientific libraries,
to find out the state-of-the-art in the three approaches mentioned above.
The results show that dengue modeling is constantly growing.
The most frequent diagnostic models were based on LoR. LoR is one
of the most used modeling techniques because of its ease of realization
and interpretation of results. Although other techniques, such as deci­
sion trees, can be easily interpreted, they consist of a large number of
nodes, which can require a significant amount of mental effort to un­
derstand a particular prediction. In contrast, an LoR model is simply a
list of coefficients, which is attractive to know the influence of charac­
teristics on the target variable. In addition, regression does not require
that the continuous independent variables follow a normal distribution,
and continuous and discrete predictors can be used together. With
respect to the category of features, the most used for these models were
demographic, clinical and laboratory data. Data are readily available
from local health authorities in each country.
In general, the most frequent epidemic models were based on LiR, RF
and SVM, with socioeconomic, demographic, climatic and environ­
mental data. From this category, the most explored approach is the
spatial-temporal analysis of dengue and its transmission vector. These
techniques are commonly implemented to map disease risk in endemic
areas, and establish relationships between risk factors and dengue
incidence.
Studies on intervention systems for dengue are quite limited. In this
review, we found only eight studies with developed models for disease
control. The techniques used were GLM, SS y BTS. The main data used
were entomological. The morbidity and mortality of the disease clearly
depend on the decisions made by health authorities, therefore, more
studies are needed in this field to support the decisions. Finally, an
important general remark is that the diagnostic, epidemic and inter­
vention models of dengue are normally machine learning models of
predictive, diagnostic or prescriptive type.
Several limitations were found in the reviewed papers, among which
we have: the absence of reporting of preprocessing techniques used, and
small sample sizes for disease variations, such as SD and DSS. Reviewing
the strengths and limitations of the articles allowed the identification of
future works for research: i) cause-effect models for dengue diagnosis, ii)
use of new features, such as genetic data and Raman spectroscopy, for
disease diagnosis, iii) a preprocessing phase based on feature engi­
neering processes, iv) implementation of Bayesian or fuzzy models that
adequately manage data uncertainty, v) automatic prescriptive models
for data-based decision-making systems, and vi) models combining the
three approaches discussed in this article using autonomous cycles of
data analysis.
Based on these future works, prioritization should focus on cause-
effect models for disease diagnosis. Not only the detection of the dis­
ease is critical, but also, the assessment of factors that most influence the
infection. A better understanding of dengue-related causes with more
13
Artificial Intelligence In Medicine 119 (2021) 102157
robust diagnostic models, would help considerably in prevention and
reducing complications and deaths. Likewise, modeling for the man­
agement of data uncertainty is urgent. The low quality of epidemio­
logical data on dengue is one of the main obstacles for the improvement
of existing models. The use of other types of predictors, such as genetic
data and spectrograms, could be useful, but their high cost of determi­
nation and collection could be a limitation. Finally, the use of autono­
mous cycles of data analysis tasks would automate decisions for disease
control.
This study has a few limitations. The first limitation is that some
online databases (ScienceDirect, IEEE Xplorer, Google Scholar, Emerald,
Taylor & Francis and PubMed) were used, and interesting articles from
other digital libraries could have been ignored. Second, the language
chosen was English because most of the latest advances in dengue
modeling are written and published in this language. The absence of
articles in other languages, such as Spanish and Portuguese, limits the
scope of the results of this study.
CRediT authorship contribution statement
All authors have participated in the design of the original study, data
analysis, interpretation of results. They also participated in the writing,
review and approval of the final manuscript.
Declaration of competing interest
The authors declare no conflict of interest.
Acknowledgements
This study was partially funded by Colombian Administrative
Department of Science, Technology and Innovation - COLCIENCIAS
(grant number 111572553478) (M. Toro) and Colombian Ministry of
Science and Technology Bicentennial PhD Grant (W. Hoyos).
References
[1] Villar LA, Rojas DP, Besada-Lombana S, Sarti E. Epidemiological trends of dengue
disease in Colombia (2000− 2011): a systematic review. PLoS Negl Trop Dis 2015;
9:e0003499. https://doi.org/10.1371/journal.pntd.0003499.
[2] Savargaonkar D, Sinha S, Srivastava B, Nagpal BN, Sinha A, Shamim A, et al. An
epidemiological study of dengue and its coinfections in Delhi. Int J Infect Dis
2018;74:41–6. https://doi.org/10.1016/j.ijid.2018.06.020.
[3] Martina B, Koraka P, Osterhaus A. Dengue virus pathogenesis: an integrated view.
2009. https://doi.org/10.1128/CMR.00035-09.
[4] Wilder-Smith A, Ooi E-E, Horstick O, Wills B. Dengue. Lancet 2019;393:350–63.
https://doi.org/10.1016/S0140-6736(18)32560-1.
[5] World Health Organization. Dengue hemorrhagic fever: diagnosis, treatment,
prevention and control. In: Technical report. WHO; 1997. URL: http://www.who.
int/csr/resources/publications/dengue/Denguepublication/en (Accessed 20
April 2020).
[6] World Health Organization. Dengue guidelines for diagnosis, treatment,
prevention and control: new edition. In: Technical report. WHO; 2009. URL: https
://apps.who.int/iris/bitstream/handle/10665/44188/9789241547871_eng.pdf?
sequence=1&isAllowed=y (Accessed 10 April 2020).
[7] World Health Organization. Dengue and severe dengue. In: Technical report.
WHO; 2010. URL: https://www.who.int/denguecontrol/en/ (Accessed 12 June
2020).
[8] Jing Q, Wang M. Dengue epidemiology. Glob Health J 2019;3:37–45. https://doi.
org/10.1016/j.glohj.2019.06.002.
[9] Alhaeli A, Bahkali S, Ali A, Househ MS, El-Metwally AA. The epidemiology of
dengue fever in Saudi Arabia: a systematic review. 2016. https://doi.org/
10.1016/j.jiph.2015.05.006.
[10] Humphrey JM, Cleton NB, Reusken CB, Glesby MJ, Koopmans MP, Abu-
Raddad LJ. Dengue in the Middle East and North Africa: a systematic review.
PLoS Negl Trop Dis 2016;10. https://doi.org/10.1371/journal.pntd.0005194.
[11] Limkittikul K, Brett J, L’Azou M. Epidemiological trends of dengue disease in
Thailand (2000–2011): a systematic literature review. PLoS Negl Trop Dis 2014;
8. https://doi.org/10.1371/journal.pntd.0003241.
[12] Mohd-Zaki AH, Brett J, Ismail E, L’Azou M. Epidemiology of dengue disease in
Malaysia (2000− 2012): a systematic literature review. PLoS Negl Trop Dis 2014;
8. https://doi.org/10.1371/journal.pntd.0003159.
[13] Bravo L, Roque VG, Brett J, Dizon R, L’Azou M. Epidemiology of dengue disease
in the Philippines (2000–2011): a systematic literature review. PLoS Negl Trop
Dis 2014;8. https://doi.org/10.1371/journal.pntd.0003027.
W. Hoyos et al.
[14] Dantés HG, Farfán-Ale JA, Sarti E. Epidemiological trends of dengue disease in
Mexico (2000–2011): a systematic literature search and analysis. PLoS Negl Trop
Dis 2014;8. https://doi.org/10.1371/journal.pntd.0003158.
[15] Teixeira MG, Siqueira JB, Ferreira GL, Bricks L, Joint G. Epidemiological trends of
dengue disease in Brazil (2000− 2010): a systematic literature search and
analysis. 2013. https://doi.org/10.1371/journal.pntd.0002520.
[16] Lim JK, Alexander N, Di Tanna GL. A systematic review of the economic impact of
rapid diagnostic tests for dengue. BMC Health Serv Res 2017;17. https://doi.org/
10.1186/s12913-017-2789-8.
[17] Luo R, Fongwen N, Kelly-Cirino C, Harris E, Wilder-Smith A, Peeling RW. Rapid
diagnostic tests for determining dengue serostatus: a systematic review and key
informant interviews. 2019. https://doi.org/10.1016/j.cmi.2019.01.002.
[18] Endo IC, Ziegelmann PK, Patel A. The economic promise of developing and
implementing dengue vaccines: evidence from a systematic review. 2016.
https://doi.org/10.1016/j.vaccine.2016.10.037.
[19] Supadmi W, Suwantika AA, Perwitasari DA, Abdulah R. Economic evaluations of
dengue vaccination in Southeast Asia Region: evidence from a systematic review.
In: Value in health regional issues. vol. 18; 2019. p. 132–44. https://doi.org/
10.1016/j.vhri.2019.02.004.
[20] Agarwal R, Wahid MH, Yausep OE, Angel SH, Lokeswara AW. The
immunogenicity and safety of CYD-Tetravalent Dengue Vaccine (CYD-TDV) in
children and adolescents: a systematic review. Acta Med Indones 2017;49(1). htt
ps://pubmed.ncbi.nlm.nih.gov/28450651/.
[21] Da Silveira LTC, Tura B, Santos M. Systematic review of dengue vaccine efficacy.
2019. https://doi.org/10.1186/s12879-019-4369-5.
[22] Godói IP, Lemos LLP, Araújo VE De, Bonoto BC, Godman B, Júnior AA Guerra.
CYD-TDV dengue vaccine: systematic review and meta-analysis of efficacy,
immunogenicity and safety. 2017. https://doi.org/10.2217/cer-2016-0045.
[23] Louis VR, Phalkey R, Horstick O, Ratanawong P, Wilder-Smith A, Tozan Y, et al.
Modeling tools for dengue risk mapping - a systematic review. Int J Health Geogr
2014;13. https://doi.org/10.1186/1476-072X-13-50.
[24] Naish S, Dale P, Mackenzie JS, McBride J, Mengersen K, Tong S. Climate change
and dengue: a critical and systematic review of quantitative modeling
approaches. BMC Infect Dis 2014;14. https://doi.org/10.1186/1471-2334-14-
167.
[25] Andraud M, Hens N, Marais C, Beutels P. Dynamic epidemiological models for
dengue transmission: a systematic review of structural approaches. PLoS One
2012;7. https://doi.org/10.1371/journal.pone.0049085.
[26] Lourenço J, Tennant W, Faria NR, Walker A, Gupta S, Recker M. Challenges in
dengue research: a computational perspective. Evol Appl 2018;11:516–33.
https://doi.org/10.1111/eva.12554.
[27] Moher D, Liberati A, Tetzlaff J, Altman DG, Altman D, Antes G, et al. Preferred
reporting items for systematic reviews and meta-analyses: the PRISMA statement.
PLoS Med 2009;6:e1000097. https://doi.org/10.1371/journal.pmed.1000097.
[28] OECD, Gross domestic spending on R&D. URL: https://data.oecd.org/rd/gross-do
mestic-spending-on-r-d.htm; 2019 (Accessed 30 April 2020).
[29] Macedo-Hair G, Nobre F Fonseca, Brasil P. Characterization of clinical patterns of
dengue patients using an unsupervised machine learning approach. BMC Infect
Dis 2019;19:1–11. https://doi.org/10.1186/s12879-019-4282-y.
[30] Fernández E, Smieja M, Walter SD, Loeb M. A predictive model to differentiate
dengue from other febrile illness. BMC Infect Dis 2016;16:1–7. https://doi.org/
10.1186/s12879-016-2024-y.
[31] Chatterjee S, Dey N, Shi F, Ashour AS, Fong SJ, Sen S. Clinical application of
modified bag-of-features coupled with hybrid neural-based classifier in dengue
fever classification using gene expression data. Med Biol Eng Comput 2018;56:
709–20. https://doi.org/10.1007/s11517-017-1722-y.
[32] Gambhir S, Malik SK, Kumar Y. PSO-ANN based diagnostic model for the early
detection of dengue disease. New Horizons Transl Med 2017;4:1–8. https://doi.
org/10.1016/j.nhtm.2017.10.001.
[33] J. Schmidhuber, Deep learning in neural networks: an overview, Neural Netw 61
(2015) 85–117. doi: https://doi.org/10.1016/j.neunet.2014.09.003. arXiv:
1404.7828.
[34] Ho TS, Weng TC, Wang JD, Han HC, Cheng HC, Yang CC, et al. Comparing
machine learning with case-control models to identify confirmed dengue cases.
PLoS Negl Trop Dis 2020;14:1–21. https://doi.org/10.1371/journal.
pntd.0008843.
[35] Park S, Srikiatkhachorn A, Kalayanarooj S, Macareo L, Green S, Friedman JF,
et al. Use of structural equation models to predict dengue illness phenotype. PLoS
Negl Trop Dis 2018;12:e0006799. https://doi.org/10.1371/journal.
pntd.0006799.
[36] Khosavanna RR, Kareko BW, Brady AC, Booty BL, Nix CD, Lyski ZL, et al. Clinical
symptoms of dengue infection among patients from a non-endemic area and
potential for a predictive model: A multiple logistic regression analysis and
decision tree. Am J Trop Med Hyg 2021;104:121–9. https://doi.org/10.4269/
AJTMH.20-0192.
[37] Eick SM, Dale AP, McKay B, Lawrence C, Ebell MH, Cordero JF, et al.
Seroprevalence of Dengue and Zika Virus in blood donations: a systematic review.
Transfus Med Rev 2019;33:35–42. https://doi.org/10.1016/j.tmrv.2018.10.001.
[38] Al-Raddadi R, Alwafi O, Shabouni O, Akbar N, Alkhalawi M, Ibrahim A, et al.
Seroprevalence of dengue fever and the associated sociodemographic, clinical,
and environmental factors in Makkah, Madinah, Jeddah, and Jizan, Kingdom of
Saudi Arabia. Acta Trop 2019;189:54–64. https://doi.org/10.1016/j.
actatropica.2018.09.009.
[39] Aguas R, Dorigatti I, Coudeville L, Luxemburger C, Ferguson NM. Cross-serotype
interactions and disease outcome prediction of dengue infections in Vietnam. Sci
Rep 2019;9:1–12. https://doi.org/10.1038/s41598-019-45816-6.
14
Artificial Intelligence In Medicine 119 (2021) 102157
[40] Jayasundara SD, Perera SS, Malavige GN, Jayasinghe S. Mathematical modeling
and a systems science approach to describe the role of cytokines in the evolution
of severe dengue. BMC Syst Biol 2017;11:1–14. https://doi.org/10.1186/s12918-
017-0415-3.
[41] Low GKK, Gan SC, Zainal N, Naidu KD, Amin-Nordin S, Khoo CS, et al. The
predictive and diagnostic accuracy of vascular endothelial growth factor and
pentraxin-3 in severe dengue. In: Pathogens and Global Health. vol. 112; 2018.
p. 334–41. https://doi.org/10.1080/20477724.2018.1516417.
[42] Nargis HF, Nawaz H, Ditta A, Mahmood T, Majeed MI, Rashid N, et al. Raman
spectroscopy of blood plasma samples from breast cancer patients at different
stages. Spectrochim Acta A Mol Biomol Spectrosc 2019;222:117210. https://doi.
org/10.1016/j.saa.2019.117210.
[43] Bahreini M, Hosseinzadegan A, Rashidi A, Miri SR, Mirzaei HR, Hajian P.
A Raman-based serum constituents’ analysis for gastric cancer diagnosis: in vitro
study, Talantavol. 204; 2019. p. 826–32. https://doi.org/10.1016/j.
talanta.2019.06.068.
[44] Shao L, Zhang A, Rong Z, Wang C, Jia X, Zhang K, et al. Fast and non-invasive
serum detection technology based on surface-enhanced Raman spectroscopy and
multivariate statistical analysis for liver disease. Nanomedicine 2018;14:451–9.
https://doi.org/10.1016/j.nano.2017.11.022.
[45] Gurian E, Giraudi P, Rosso N, Tiribelli C, Bonazza D, Zanconati F, et al.
Differentiation between stages of non-alcoholic fatty liver diseases using surface-
enhanced Raman spectroscopy. Anal Chim Acta 2020;1110:190–8. https://doi.
org/10.1016/j.aca.2020.02.040.
[46] Khan S, Ullah R, Shahzad S, Anbreen N, Bilal M, Khan A. Analysis of tuberculosis
disease through Raman spectroscopy and machine learning. Photodiagn
Photodyn Ther 2018;24:286–91. https://doi.org/10.1016/j.pdpdt.2018.10.014.
[47] Pérez A, Prada YA, Cabanzo R, González CI, Mejía-Ospino E. Diagnosis of chagas
disease from human blood serum using surface-enhanced Raman scattering
(SERS) spectroscopy and chemometric methods. Sens Bio-Sens Res 2018;21:40–5.
https://doi.org/10.1016/j.sbsr.2018.10.003.
[48] Khan S, Ullah R, Khan A, Sohail A, Wahab N, Bilal M, et al. Random Forest-based
evaluation of Raman spectroscopy for dengue fever analysis. Appl Spectrosc
2017;71:2111–7. https://doi.org/10.1177/0003702817695571.
[49] Khan S, Ullah R, Khan A, Wahab N, Bilal M, Ahmed M. Analysis of dengue
infection based on Raman spectroscopy and support vector machine (SVM).
Biomed Opt Express 2016;7:2249. https://doi.org/10.1364/boe.7.002249.
[50] Suwarto S, Nainggolan L, Sinto R, Effendi B, Ibrahim E, Suryamin M, et al.
Dengue score: a proposed diagnostic predictor for pleural effusion and/or ascites
in adults with dengue infection. BMC Infect Dis 2016;16:1–7. https://doi.org/
10.1186/s12879-016-1671-3.
[51] Silva NS da, Undurraga EA, Silva Ferreira ER da, Estofolete CF, Nogueira ML.
Clinical, laboratory, and demographic determinants of hospitalization due to
dengue in 7613 patients: a retrospective study based on hierarchical models. Acta
Trop 2018;177:25–31. https://doi.org/10.1016/j.actatropica.2017.09.025.
[52] Fernández E, Smieja M, Walter SD, Loeb M. A retrospective cohort study to
predict severe dengue in Honduran patients. BMC Infect Dis 2017;17. https://doi.
org/10.1186/s12879-017-2800-3.
[53] Phuong NTN, Manh DH, Dumre SP, Mizukami S, Weiss LN, Thuong N Van, et al.
Plasma cell-free DNA: a potential biomarker for early prediction of severe dengue.
Ann Clin Microbiol Antimicrob 2019;18. https://doi.org/10.1186/s12941-019-
0309-x.
[54] Davi CCM, Pastor A, Oliveira T, Neto FB Lima, Braga-Neto U, Bigham A, et al.
Severe dengue prognosis using human genome data and machine learning. IEEE
Trans Biomed Eng 2019. https://doi.org/10.1109/TBME.2019.2897285.
[55] Tuan NM, Nhan HT, Van Vinh Chau N, Hung NT, Tuan HM, Tram T Van, et al. An
evidence-based algorithm for early prognosis of severe dengue in the outpatient
setting. Clin Infect Dis 2017;64:656–63. https://doi.org/10.1093/cid/ciw863.
[56] Ahmad MH, Ibrahim MI, Mohamed Z, Ismail N, Abdullah MA, Shueb RH, et al.
The sensitivity, specificity and accuracy of warning signs in predicting severe
dengue, the severe dengue prevalence and its associated factors. Int J Environ Res
Public Health 2018;15. https://doi.org/10.3390/ijerph15092018.
[57] Phakhounthong K, Chaovalit P, Jittamala P, Blacksell SD, Carter MJ, Turner P,
et al. Predicting the severity of dengue fever in children on admission based on
clinical features and laboratory indicators: application of classification tree
analysis. BMC Pediatr 2018;18:1–9. https://doi.org/10.1186/s12887-018-1078-
y.
[58] Huang SW, Tsai HP, Hung SJ, Ko WC, Wang JR. Assessing the risk of dengue
severity using demographic information and laboratory test results with machine
learning. PLoS Negl Trop Dis 2020;14:1–19. https://doi.org/10.1371/journal.
pntd.0008960.
[59] Zhang H, Xie Z, Xie X, Ou Y, Zeng W, Zhou Y. A novel predictor of severe dengue:
the aspartate aminotransferase/platelet count ratio index (APRI). J Med Virol
2018;90:803–9. https://doi.org/10.1002/jmv.25021.
[60] Lin CY, Kolliopoulos C, Huang CH, Tenhunen J, Heldin CH, Chen YH, et al. High
levels of serum hyaluronan is an early predictor of dengue warning signs and
perturbs vascular integrity. EBioMedicine 2019;48:425–41. https://doi.org/
10.1016/j.ebiom.2019.09.014.
[61] Lee IK, Liu JW, Chen YH, Chen YC, Tsai CY, Huang SY, et al. Development of a
simple clinical risk score for early prediction of severe dengue in adult patients.
PLoS One 2016;11:e0154772. https://doi.org/10.1371/journal.pone.0154772.
[62] Lam PK, Tam DTH, Dung NM, Tien NTH, Kieu NTT, Simmons C, et al.
A prognostic model for development of profound shock among children
presenting with dengue shock syndrome. PLoS One 2015;10:1–13. https://doi.
org/10.1371/journal.pone.0126134.
W. Hoyos et al.
[63] Lam PK, Ngoc TV, Thuy TT Thu, Van NT Hong, Thuy TT Nhu, Tam DT Hoai, et al.
The value of daily platelet counts for predicting dengue shock syndrome: results
from a prospective observational study of 2301 Vietnamese children with dengue.
PLoS Negl Trop Dis 2017;11. https://doi.org/10.1371/journal.pntd.0005498.
[64] Rossi G, Karki S, Smith RL, Brown WM, Ruiz MO. The spread of mosquito-borne
viruses in modern times: A spatio-temporal analysis of dengue and chikungunya.
Spatial Spatio-temporal Epidemiol 2018;26:113–25. https://doi.org/10.1016/j.
sste.2018.06.002.
[65] Delmelle E, Hagenlocher M, Kienberger S, Casas I. A spatial model of
socioeconomic and environmental determinants of dengue fever in Cali,
Colombia. Acta Trop 2016;164:169–76. https://doi.org/10.1016/j.
actatropica.2016.08.028.
[66] Mao L, Yin L, Song X, Mei S. Mapping intra-urban transmission risk of dengue
fever with big hourly cellphone data. Acta Trop 2016;162:188–95. https://doi.
org/10.1016/j.actatropica.2016.06.029.
[67] Mutheneni SR, Mopuri R, Naish S, Gunti D, Upadhyayula SM. Spatial distribution
and cluster analysis of dengue using self organizing maps in Andhra Pradesh,
India, 2011–2013. Parasite Epidemiol Contr 2018;3:52–61. https://doi.org/
10.1016/j.parepi.2016.11.001.
[68] Akter R, Naish S, Hu W, Tong S. Socio-demographic, ecological factors and
dengue infection trends in Australia. PLoS One 2017;12:1–18. https://doi.org/
10.1371/journal.pone.0185551.
[69] Y. Yue, J. Sun, X. Liu, D. Ren, Q. Liu, X. Xiao, L. Lu, Spatial analysis of dengue
fever and exploration of its environmental and socio-economic risk factors using
ordinary least squares: a case study in five districts of Guangzhou City, China,
2014, Int J Infect Dis 75 (2018) 39–48. doi: https://doi.org/10.1016/j.ijid.2018.0
7.023.
[70] Reyes-Castro PA, Harris RB, Brown HE, Christopherson GL, Ernst KC. Spatio-
temporal and neighborhood characteristics of two dengue outbreaks in two arid
cities of Mexico. Acta Trop 2017;167:174–82. https://doi.org/10.1016/j.
actatropica.2017.01.001.
[71] Ministry of Health. Epidemiological report–dengue fever (January to June,
2008)., Technical Report, Ministry of Health. URL: http://bvsms.saude.gov.br
/bvs/publicacoes/informe_epidemiologico_dengue_janeiro_junho_2008.pdf;
2008.
[72] Stolerman LM, Maia PD, Kutz JN. Forecasting dengue fever in Brazil: an
assessment of climate conditions. PLoS One 2019;14:e0220106. https://doi.org/
10.1371/journal.pone.0220106.
[73] Carvajal TM, Viacrusis KM, Hernandez LFT, Ho HT, Amalin DM, Watanabe K.
Machine learning methods reveal the temporal pattern of dengue incidence using
meteorological factors in metropolitan Manila, Philippines. BMC Infect Dis 2018;
18:1–15. https://doi.org/10.1186/s12879-018-3066-0.
[74] Zhao N, Charland K, Carabali M, Nsoesie EO, Maheu-Giroux M, Rees E, et al.
Machine learning and dengue forecasting: comparing random forests and
artificial neural networks for predicting dengue burden at national and sub-
national scales in Colombia. PLoS Negl Trop Dis 2020;14:1–16. https://doi.org/
10.1371/journal.pntd.0008056.
[75] Salim NAM, Wah YB, Reeves C, Smith M, Yaacob WFW, Mudin RN, et al.
Prediction of dengue outbreak in Selangor Malaysia using machine learning
techniques. Sci Rep 2021;11. https://doi.org/10.1038/s41598-020-79193-2.
[76] Salami D, Sousa CA, R. O. Martins Md, Capinha C. Predicting dengue importation
into Europe, using machine learning and model-agnostic methods. Sci Rep 2020;
10. https://doi.org/10.1038/s41598-020-66650-1.
[77] Chang FS, Tseng YT, Hsu PS, Chen CD, Lian IB, Chao DY. Re-assess vector indices
threshold as an early warning tool for predicting dengue epidemic in a dengue
non-endemic country. PLoS Negl Trop Dis 2015;9. https://doi.org/10.1371/
journal.pntd.0004043.
[78] J. M. Scavuzzo, F. Trucco, M. Espinosa, C. B. Tauro, M. Abril, C. M. Scavuzzo, A.
C. Frery, Modeling dengue vector population using remotely sensed data and
machine learning, Acta Trop 185 (2018) 167–175. doi: https://doi.org/10.101
6/j.actatropica.2018.05.003. arXiv:1805.02590.
[79] Parra MCP, Fávaro EA, Dibo MR, Mondini A, Eiras ÁE, Kroon EG, et al. Using
adult Aedes aegypti females to predict areas at risk for dengue transmission: a
spatial case-control study. Acta Trop 2018;182:43–53. https://doi.org/10.1016/j.
actatropica.2018.02.018.
[80] Ding F, Fu J, Jiang D, Hao M, Lin G. Mapping the spatial distribution of Aedes
aegypti and Aedes albopictus. Acta Trop 2018;178:155–62. https://doi.org/
10.1016/j.actatropica.2017.11.020.
[81] Jácome G, Vilela P, Yoo CK. Present and future incidence of dengue fever in
Ecuador nationwide and coast region scale using species distribution modeling
for climate variability’s effect. Ecol Model 2019;400:60–72. https://doi.org/
10.1016/j.ecolmodel.2019.03.014.
[82] Wu CH, Kao SC, Shih CH, Kan MH. Open data mining for Taiwan’s dengue
epidemic. Acta Trop 2018;183:1–7. https://doi.org/10.1016/j.
actatropica.2018.03.017.
[83] Strauss RA, Castro JS, Reintjes R, Torres JR. Google dengue trends: an indicator of
epidemic behavior. Venezuelan Case Int J Med Informatics 2017;104:26–30.
https://doi.org/10.1016/j.ijmedinf.2017.05.003.
[84] Li Z, Liu T, Zhu G, Lin H, Zhang Y, He J, et al. Dengue Baidu Search Index data
can improve the prediction of local dengue epidemic: a case study in Guangzhou,
China. PLoS Negl Trop Dis 2017;11:e0005354. https://doi.org/10.1371/journal.
pntd.0005354.
[85] Liu K, Wang T, Yang Z, Huang X, Milinovich GJ, Lu Y, et al. Using Baidu Search
Index to predict dengue outbreak in China. Sci Rep 2016;6. https://doi.org/
10.1038/srep38040.
15
Artificial Intelligence In Medicine 119 (2021) 102157
[86] d. A. Marques-Toledo C, Degener CM, Vinhal L, Coelho G, Meira W, Codeço CT,
et al. Dengue prediction by the web: tweets are a useful tool for estimating and
forecasting Dengue at country and city level. PLoS Negl Trop Dis 2017;11.
https://doi.org/10.1371/journal.pntd.0005729.
[87] Ramadona AL, Tozan Y, Lazuardi L, Rocklöv J. A combination of incidence data
and mobility proxies from social media predicts the intraurban spread of dengue
in Yogyakarta, Indonesia. PLoS Negl Trop Dis 2019;13. https://doi.org/10.1371/
journal.pntd.0007298.
[88] Souza RC, Assunção RM, Oliveira DM, Neill DB, Meira W. Where did I get
dengue? Detecting spatial clusters of infection risk with social network data.
Spatial Spatio-temporal Epidemiol 2019;29:163–75. https://doi.org/10.1016/j.
sste.2018.11.005.
[89] C. C. Huang, C. C. Hsu, H. R. Guo, S. B. Su, H. J. Lin, Dengue fever mortality score:
A novel decision rule to predict death from dengue fever, J Inf Secur 75 (2017 a)
532–540. doi: https://doi.org/10.1016/j.jinf.2017.09.014.
[90] Huang HS, Hsu CC, Ye JC, Su SB, Huang CC, Lin HJ. Predicting the mortality in
geriatric patients with dengue fever. Medicine 2017 b;96. https://doi.org/
10.1097/MD.0000000000007878.
[91] Kesorn K, Ongruk P, Chompoosri J, Phumee A, Thavara U, Tawatsin A, et al.
Morbidity rate prediction of dengue hemorrhagic fever (DHF) using the support
vector machine and the Aedes aegypti infection rate in similar climates and
geographical areas. PLoS One 2015;10:1–16. https://doi.org/10.1371/journal.
pone.0125049.
[92] Md-Sani SS, Md-Noor J, Han WH, Gan SP, Rani NS, Tan HL, et al. Prediction of
mortality in severe dengue cases. BMC Infect Dis 2018;18. https://doi.org/
10.1186/s12879-018-3141-6.
[93] Nagori A, Dhingra LS, Bhatnagar A, Lodha R, Sethi T. Predicting hemodynamic
shock from thermal images using machine learning. Sci Rep 2019;9. https://doi.
org/10.1038/s41598-018-36586-8.
[94] Dua X, Rajendra A. Data mining in biomedical imaging, signaling, and systems.
Boca Raton: Auerbach Publications; 2016. https://doi.org/10.1201/b10917.
[95] Kalimuthu K, Panneerselvam C, Chou C, Tseng LC, Murugan K, Tsai KH, et al.
Control of dengue and Zika virus vector Aedes aegypti using the predatory
copepod Megacyclops formosanus: synergy with Hedychium coronarium-
synthesized silver nanoparticles and related histological changes in targeted
mosquitoes. Process Saf Environ Prot 2017;109:82–96. https://doi.org/10.1016/
j.psep.2017.03.027.
[96] Udayanga L, Ranathunge T, Iqbal MC, Abeyewickreme W, Hapugoda M.
Predatory efficacy of five locally available copepods on Aedes larvae under
laboratory settings: an approach towards bio-control of dengue in Sri Lanka. PLoS
One 2019;14:1–14. https://doi.org/10.1371/journal.pone.0216140.
[97] Lee SJ, Kim S, Yu JS, Kim JC, Nai YS, Kim JS. Biological control of Asian tiger
mosquito, Aedes albopictus (Diptera: Culicidae) using Metarhizium anisopliae
JEF-003 millet grain. J Asia Pac Entomol 2015;18:217–21. https://doi.org/
10.1016/j.aspen.2015.02.003.
[98] Benelli G, Jeffries CL, Walker T. Biological control of mosquito vectors: past,
present, and future. Insects 2016;7:1–18. https://doi.org/10.3390/
insects7040052.
[99] Nazni WA, Hoffmann AA, NoorAfizah A, Cheong YL, Mancini MV, Golding N,
et al. Establishment of Wolbachia Strain wAlbB in Malaysian populations of Aedes
aegypti for dengue control. Curr Biol 2019;29. https://doi.org/10.1016/j.
cub.2019.11.007 (4241–4248.e5).
[100] Indriani C, Tantowijoyo W, Rancès E, Andari B, Prabowo E, Yusdi D, et al.
Reduced dengue incidence following deployments of Wolbachia-infected Aedes
aegypti in Yogyakarta, Indonesia: a quasi-experimental trial using controlled
interrupted time series analysis. Gates Open Res 2020;4:1–16. https://doi.org/
10.12688/gatesopenres.13122.1.
[101] Ryan PA, Turley AP, Wilson G, Hurst TP, Retzki K, Brown-Kenyon J, et al.
Establishment of wMel Wolbachia in Aedes aegypti mosquitoes and reduction of
local dengue transmission in Cairns and surrounding locations in northern
Queensland, Australia. Gates Open Res 2020;3:1547. https://doi.org/10.12688/
gatesopenres.13061.2.
[102] Lee JS, Lim JK, Dang DA, Nguyen THA, Farlow A. Dengue vaccine supplies under
endemic and epidemic conditions in three dengue-endemic countries: Colombia,
Thailand, and Vietnam. Vaccine 2017;35:6957–66. https://doi.org/10.1016/j.
vaccine.2017.10.070.
[103] Hladish TJ, Pearson CA, Rojas D Patricia, Gomez-Dantes H, Halloran ME,
Vazquez-Prokopec GM, et al. Forecasting the effectiveness of indoor residual
spraying for reducing dengue burden. PLoS Negl Trop Dis 2018;12:1–16. https://
doi.org/10.1371/journal.pntd.0006570.
[104] Salimi A, Ziaii M, Amiri A, Zadeh M Hosseinjani, Karimpouli S, Moradkhani M.
Using a Feature Subset Selection method and Support Vector Machine to address
curse of dimensionality and redundancy in Hyperion hyperspectral data
classification. Egypt J Remote Sens Space Sci 2018;21:27–36. https://doi.org/
10.1016/j.ejrs.2017.02.003.
[105] Araujo M, Aguilar J, Aponte H. Fault detection system in gas lift well based on
artificial immune system. In: Proceedings of the international joint conference on
neural Networks. vol. 3. IEEE; 2003. p. 1673–7. https://doi.org/10.1109/
ijcnn.2003.1223658.
[106] Puerto E, Aguilar J, Vargas R, Reyes J. An ar2p deep learning architecture for the
discovery and the selection of features. Neural Process Lett 2019;50:623–43.
https://doi.org/10.1007/s11063-019-10062-4.
[107] Li Y, Sperrin M, Martin GP, Ashcroft DM, van Staa TP. Examining the impact of
data quality and completeness of electronic health records on predictions of
patients’ risks of cardiovascular disease. Int J Med Inform 2020;133:104033.
https://doi.org/10.1016/j.ijmedinf.2019.104033.
W. Hoyos et al.
[108] Chen L, Duan G, Wang SY, Ma JF. A Choquet integral based fuzzy logic approach
to solve uncertain multi-criteria decision making problem. Expert Syst Appl 2020;
149:113303. https://doi.org/10.1016/j.eswa.2020.113303.
[109] Qin B, Xia Y, Wang S, Du X. A novel Bayesian classification for uncertain data.
Knowl-Based Syst 2011;24:1151–8. https://doi.org/10.1016/j.
knosys.2011.04.011.
[110] Shaheen O, El-Nagar AM, El-Bardini M, El-Rabaie NM. Probabilistic fuzzy logic
controller for uncertain nonlinear systems. J Frankl Inst 2018;355:1088–106.
https://doi.org/10.1016/j.jfranklin.2017.12.015.
[111] Sugasawa S, Kubokawa T. Bayesian estimators in uncertain nested error
regression models. J Multivar Anal 2017;153:52–63. https://doi.org/10.1016/j.
jmva.2016.09.011.
[112] Velasco H, Laniado H, Toro M, Leiva V, Lio Y. Robust three-step regression based
on comedian and its performance in cell-wise and case-wise outliers. Mathematics
2020;8:1259. https://doi.org/10.3390/math8081259.
[113] Quick H, Waller LA. Using spatiotemporal models to generate synthetic data for
public use. Spatial Spatio-temporal Epidemiol 2018;27:37–45. https://doi.org/
10.1016/j.sste.2018.08.004.
16
Artificial Intelligence In Medicine 119 (2021) 102157
[114] Silver M, Karnieli A, Marra F, Fredj E. An evaluation of weather radar adjustment
algorithms using synthetic data. J Hydrol 2019;576:408–21. https://doi.org/
10.1016/j.jhydrol.2019.06.064.
[115] Lepenioti K, Bousdekis A, Apostolou D, Mentzas G. Prescriptive analytics:
literature review and research challenges. Int J Inf Manag 2020;50:57–70.
https://doi.org/10.1016/J.IJINFOMGT.2019.04.003.
[116] Menezes BC, Kelly JD, Leal AG, Roux GC Le. Predictive, prescriptive and detective
analytics for smart manufacturing in the information age. IFAC-PapersOnLine
2019;52:568–73. https://doi.org/10.1016/J.IFACOL.2019.06.123.
[117] Google. Federated learning: collaborative machine learning without centralized
training data. URL: https://ai.googleblog.com/2017/04/federated-learning-colla
borative.html; 2017.
[118] Aguilar J, Sánchez M, Cordero J, Valdiviezo-Díaz P, Barba-Guamán L, Chamba-
Eras L. Learning analytics tasks as services in smart classrooms. Univ Access Inf
Soc 2018;17:693–709. https://doi.org/10.1007/s10209-017-0525-0.
[119] Aguilar J, Cordero J, Buendia O. Specification of the autonomic cycles of learning
analytic tasks for a smart classroom. J Educ Comput Res 2018;56:866–91. https://
doi.org/10.1177/0735633117727698.